Kimi-K3 IQ1_S REAP568 — current 600K four-Spark profile
This repository contains the four-Spark Kimi-K3 IQ1_S derivative currently
served as kimi-k3-f16k-600k-u1024. It retains 568 of 896 routed experts per
layer using the disclosed deterministic REAP568 selection and is packaged as
fourteen GGUF shards.
Current serving profile
| Setting | Value |
|---|
| Model name | Kimi-K3 IQ1_S REAP568 — FP16-K/F16-V 600K uBatch-1024 |
| Context | 600,000 tokens (n_ctx=600064) |
| KV cache | K F16, V F16 |
| Logical / physical batch | 2,048 / 1,024 |
| Parallel slots | 1 |
| CPU threads | 16 / batch threads 20 |
| Distribution | Local CUDA plus three RPC workers over RoCE; layer split 1:1:1:1 |
| API alias | kimi-k3-f16k-600k-u1024 |
The production launcher is
recipes/launch_4x_spark_600k_f16k.sh.
The 64K uBatch-1024 launcher remains available as a lower-context portable
profile.
Current measured prefill
A unique natural-language request with
cache_prompt=false on the live
600K service measured
3,255 prompt tokens at 75.09 tok/s. A separate cold
1,769-token request measured 61.30 tok/s and 2.65 tok/s for a 16-token decode.
These are request-specific measurements; prompt length, graph shape, and cache
state materially affect throughput. See
BENCHMARKS.md for
the complete record and the FP16-K comparison.
What is included
- The complete REAP568 GGUF checkpoint and tokenizer/configuration files.
- The current 600K FP16-K/FP16-V launch recipe.
- Checksum and release-verification scripts.
- Candidate-only KDA/FlashKDA optimization notes and maintenance procedure.
Expert selection and limitations
The derivative retains 568 of the original 896 routed experts in every Kimi-K3
MoE layer. Attention, KDA, MLA, AttnRes, shared experts, embeddings, latent
projections, normalization, and output tensors are unchanged. The public
sources did not provide complete per-expert REAP saliency values, so this is a
disclosed deterministic routing proxy, not a claim of lossless pruning.
FlashKDA status
FlashKDA is not enabled in the production service. The earlier candidate did
not activate the bridge because the RPC workers were using older CUDA/RPC
libraries. Bridge-enabled libraries are staged separately on all three ranks;
the production workers and API were not interrupted. The maintenance-window
procedure and promotion gates are in
recipes/FLASHKDA_CANDIDATE_RUNBOOK.md.
Validation
1sha256sum -c MANIFEST.sha256
2python recipes/verify_release.py /path/to/Kimi-K3-UD-IQ1_S-REAP568
Report cold prefill, warm prefix reuse, decode, TTFT, and quality outputs
separately. Do not treat throughput alone as an intelligence or safety claim.
Attribution, license, and responsibility
This is an independent derivative. It is not affiliated with or endorsed by
Moonshot AI, Unsloth, llama.cpp, NVIDIA, or contributors to those projects.
Comply with the upstream Kimi-K3 checkpoint, IQ1_S conversion, and runtime
licenses and notices. This release is provided for research and evaluation;
the model may produce incorrect, biased, unsafe, or unsuitable content.
Validate outputs and use appropriate access controls. This card is not legal
advice.
Sources
